AI Daily Digest · 2026-05-22
🔬 New AI Craft
1. Indexing a Year of Video Locally on a 5-Year-Old M1 Max with Gemma 4 31B
A detailed engineering build log: the author used Claude Code to write a ~1,400-line Python pipeline, running Gemma 4 31B Q4 via LM Studio on a 2021 MacBook Pro M1 Max (64GB RAM, pushed to 50GB swap), turning a year of raw camera footage into a queryable video archive. The architecture: ffprobe → exiftool GPS → WhisperX multi-language transcription with speaker diarization → insightface face embeddings → single vision model call producing full YAML sidecar description files. Unlike your typical Agent Plan workflow, Claude Code acted as the builder while the author owned architecture and prompt design — a classic upstream-leverage play.
https://blog.simbastack.com/indexed-a-year-of-video-locally/
2. Simon Willison: The Last Six Months in LLMs in Five Minutes
A concise overview of key LLM developments over the past half-year: multimodal capabilities maturing, Agent frameworks moving from concept to production, small models (<10B) rising, and open-source ecosystems approaching parity with closed-source. Good for building a bird's-eye view.
https://simonwillison.net/2026/May/19/5-minute-llms/
🛠️ Tools & Tips
1. smallcode – Coding Agent Optimized for Small LLMs, 87% on Benchmarks with 4B Active Model
A new open-source project that compresses coding agent workflows to fit a 4B-parameter active model, achieving 87% on SWE-bench. Practical for local/edge deployment alongside your existing Agent Plan workflow.
https://github.com/Doorman11991/smallcode
2. Runtime (YC P26) – Sandboxed Coding Agents for Your Whole Team
A YC-backed control plane that lets everyone on a team run Claude Code, Codex, and Gemini CLI in sandboxed environments, triggered from Slack, Linear, or CLI — with built-in guardrails, audit logs, and enterprise integrations.
https://www.runtm.com/
3. Was My $48K GPU Server Worth It? ♻️ Re-shared from 2026-05
The author self-hosted a dedicated GPU server and shares real costs (electricity, cooling, maintenance, utilization) after months of running AI inference, batch processing, and fine-tuning. Verdict: self-hosting pays off for sustained high-utilization workloads; occasional bursts are still cheaper on API.
https://rosmine.ai/2026/05/13/was-my-48k-gpu-worth-it/
⭐ Open Source Highlights
1. html-anything — 4470⭐, local agent-driven HTML editor with 75 skills × 9 surface formats (magazine, deck, resume, etc.)
https://github.com/nexu-io/html-anything
2. mirage — Unified virtual filesystem for AI Agents, cross-platform resource management
https://github.com/strukto-ai/mirage
3. clawsweeper — AI-driven Issue/PR cleanup tool, auto-scanning repos for stale items
https://github.com/openclaw/clawsweeper
4. datawhalechina/Agent-Learning-Hub — 475⭐, curated AI Agent learning roadmap
https://github.com/datawhalechina/Agent-Learning-Hub
📰 Industry News
1. Karpathy joins Anthropic, reshaping the AI talent landscape
2. Mistral AI acquires Emmi AI, accelerating European AI model-layer consolidation
3. OpenAI adopts Google's SynthID watermark for AI image provenance
🚀 Major Releases
1. Google launches Gemini 3.5 Flash with significant performance and efficiency gains
